CN114385619A - Multi-channel ocean observation time sequence scalar data missing value prediction method and system - Google Patents

Abstract

The invention belongs to the field of computer systems based on specific calculation models, and provides a method and a system for predicting missing values of multi-channel ocean observation time sequence scalar data, which are used for acquiring ocean observation time sequence scalar data with ocean missing values; obtaining a marine missing value prediction result by adopting a TA-RNN model based on the marine observation time sequence scalar data; the TA-RNN model comprises a convolution attention module, a space attention module and a time attention module, wherein the convolution attention module is used for refining the ocean observation time series scalar data; the space attention module is used for capturing dynamic space correlation of the refined ocean observation time sequence scalar data; the temporal attention module is configured to capture dynamic temporal correlations between different time intervals in the spatial attention module output data.

Description

A method and system for predicting missing values of multi-channel ocean observation time series scalar data

technical field

The invention belongs to the field of computer systems based on specific calculation models, and in particular relates to a method and system for predicting missing values of multi-channel ocean observation time series scalar data.

Background technique

The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

Ocean monitoring relies on widely deployed ocean buoys and observatories that integrate various types of ocean sensors. The complex structure of marine ecosystems makes marine observation data complex and diverse. Missing value refers to the data clustering, grouping, censoring or truncation of the original data due to lack of information, which means that one or some feature values in the data are incomplete. Because the marine observation data such as chlorophyll, wind speed, dissolved oxygen, salinity, temperature, oxygen content, wind speed, and turbidity are collected by the buoy system, navigation system and database system, each collection system is easily interfered by external environmental factors. The data has missing values. These data have implications for the accuracy of downstream applications, such as marine data assimilation and intelligent data mining. Traditional methods such as mathematical statistics and empirical prediction cannot achieve the expected goals for ocean observation data with multi-factor, irregular and complex characteristics. Therefore, data-driven research on accurate prediction models of ocean observation data plays an irreplaceable role in filling missing values in ocean observation time series scalar data.

SUMMARY OF THE INVENTION

In order to solve the technical problems existing in the above background technology, the present invention provides a method and system for predicting missing values of multi-channel ocean observation time series scalar data, which predicts its future change trend through the historical data of multi-channel ocean observation time series scalar data, The predicted data is used to fill in missing values.

In order to achieve the above object, the present invention adopts the following technical solutions:

A first aspect of the present invention provides a method for predicting missing values of multi-channel ocean observation time series scalar data.

A method for predicting missing values of multi-channel ocean observation time series scalar data, comprising:

Obtain ocean observation time series scalar data with ocean missing values;

Based on the ocean observation time series scalar data, the TA-RNN model is used to obtain the ocean missing value prediction result;

The TA-RNN model includes a convolutional attention module, a spatial attention module and a temporal attention module, the convolutional attention module is used to refine the ocean observation time series scalar data; the spatial attention module is used to capture the refinement The dynamic spatial correlation of the ocean observation time series scalar data; the time attention module is used to capture the dynamic time correlation between different time intervals in the output data of the spatial attention module.

A second aspect of the present invention provides a multi-channel ocean observation time series scalar data missing value prediction system.

A multi-channel ocean observation time series scalar data missing value prediction system, comprising:

a data acquisition module, which is configured to: acquire ocean observation time series scalar data with ocean missing values;

A prediction module, which is configured to: based on the ocean observation time series scalar data, using a TA-RNN model to obtain a prediction result of ocean missing values;

The TA-RNN model includes a convolutional attention module, a spatial attention module and a temporal attention module, the convolutional attention module is used to refine the ocean observation time series scalar data; the spatial attention module is used to capture the refinement The dynamic spatial correlation of the ocean observation time series scalar data; the time attention module is used to capture the dynamic time correlation between different time intervals in the output data of the spatial attention module.

Compared with the prior art, the beneficial effects of the present invention are:

The three-stage attention-based recurrent neural network (TA-RNN) model proposed by the present invention adopts a convolutional attention module to refine the input sequence in the first stage, so that the new input sequence has stronger representation ability; In the second stage, a spatial attention module is adopted, which enables the model to selectively capture dynamic correlations between different input sequences; in the third stage, a temporal attention module is adopted, which enables the model to adaptively capture the differences between different time intervals in the input sequence. dynamic time correlation.

The present invention can accurately fill in missing values, thereby avoiding the problems of inaccurate filling of missing values and large errors.

The invention solves the defect that the current missing value filling can only rely on single-channel data. The invention aims at ocean multi-channel observation time series scalar data, through chlorophyll and depth, temperature, conductivity, salinity, oxygen content, dissolved oxygen concentration, The correlation between ocean observation time series scalar data such as chlorophyll (including missing values), turbidity, pH value, wind speed, etc., fills in the missing values in the chlorophyll sequence. Due to the richness and variety of marine data, in most scenarios, the target sequence often does not exist alone, but coexists with many time series, which together form a specific scene dataset, and the multi-channel ocean observation time series scalar dataset is missing values. Fill, so that it is closer to the actual situation of the data set collected by the marine acquisition system.

Description of drawings

The accompanying drawings forming a part of the present invention are used to provide further understanding of the present invention, and the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, and do not constitute an improper limitation of the present invention.

1 is a flowchart of a method for predicting missing values of multi-channel ocean observation time series scalar data according to an embodiment of the present invention;

2 is a flowchart of missing value filling shown in an embodiment of the present invention;

3 is a framework diagram of a recurrent neural network model based on three-stage attention shown in an embodiment of the present invention;

FIG. 4 is a schematic diagram of a convolutional attention module (CBAM) shown in an embodiment of the present invention;

5 is a schematic diagram of a channel attention module shown in an embodiment of the present invention;

6 is a schematic diagram of a spatial attention module shown in an embodiment of the present invention;

7 is a chlorophyll sequence diagram with missing values shown in an embodiment of the present invention;

Fig. 8 is the chlorophyll sequence prediction effect diagram in the sample set with no missing value shown in the embodiment of the present invention;

Fig. 9 is the chlorophyll sequence diagram after filling shown in the embodiment of the present invention;

FIG. 10 is a chlorophyll sequence diagram with a part of the length of 50 containing missing values shown in the embodiment of the present invention;

11 is a chlorophyll sequence diagram after linear interpolation processing shown in an embodiment of the present invention;

FIG. 12 is a diagram showing the effect of filling in chlorophyll deficiency after model prediction according to an embodiment of the present invention.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and embodiments.

It should be noted that the following detailed description is exemplary and intended to provide further explanation of the invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

It should be noted that the terminology used herein is for the purpose of describing specific embodiments only, and is not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural as well, furthermore, it is to be understood that when the terms "comprising" and/or "including" are used in this specification, it indicates that There are features, steps, operations, devices, components and/or combinations thereof.

It is noted that the flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of the present disclosure. It should be noted that each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which may include one or more components used in implementing various embodiments Executable instructions for the specified logical function. It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may in fact be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It should also be noted that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented using dedicated hardware-based systems that perform the specified functions or operations , or can be implemented using a combination of dedicated hardware and computer instructions.

As described in the Background Art, the current common methods for filling missing values in marine multi-channel observation scalar data mostly use fixed values, medians and modes to fill in missing values, but this will result in inaccurate filling of missing values, and there are more problems. large errors, etc. The three-stage attention-based recurrent neural network (TA-RNN) model proposed by the present invention adopts a convolutional attention module to refine the input sequence in the first stage, so that the new input sequence has stronger representation ability; In the second stage, a spatial attention module is adopted, which enables the model to selectively capture dynamic correlations between different input sequences; in the third stage, a temporal attention module is adopted. Enables the model to adaptively capture dynamic temporal correlations between different time intervals in the input sequence. The present invention can accurately fill in missing values, thereby avoiding the problems of inaccurate filling of missing values and large errors.

The current deep learning missing value filling algorithm has the defect that it cannot fill the missing value of the multi-channel ocean observation time series scalar data. At present, the main method for filling missing values is E ² GAN, but when filling missing values with sensor input to E ² GAN, most of them only have two columns of data, time and detection value. This basically does not exist in the actual ocean scene. Multiple sensors are integrated on the ocean buoy to work at the same time, so the data collected by the ocean sensors are basically multi-channel data. Aiming at the ocean multi-channel observation time series scalar data, the invention adopts a recurrent neural network model based on three-stage attention, and uses the past value of the target sequence and the current and past values of other sequences related to the target sequence to determine the current value of the target sequence. The predicted value is filled in the missing value position of the current data set.

The present invention proposes a recurrent neural network model based on three-stage attention to accurately predict missing values of multi-channel marine data. The recurrent neural network model of three-stage attention is shown in Figure 3, wherein the three-stage attention modules are:

(1) Convolutional attention module, the convolutional attention module refines the original input sequence and increases the representation ability of the original input sequence. Among them, the convolutional attention module was proposed in 2018. It mixes spatial attention and channel attention in the convolutional module. This module is a lightweight and general module with good portability. Here we will It is used to process multi-channel input sequences.

(2) Spatial attention module, which enables the model to selectively capture dynamic spatial correlations between different input sequences.

(3) Temporal Attention Module, which enables the model to adaptively capture dynamic temporal correlations between different time intervals in the input sequence.

进行细化，生成新的输入序列

，经过卷积注意力操作后，增加了原始输入序列的表征能力；空间注意模块，它能够有选择性地捕获不同输入序列之间的动态相关性；门控循环单元，它可以学习到输入序列的隐层表示，并根据输入序列和其上一个时刻隐状态来更新当前时刻的隐状态；时间注意模块。它可以自适应地捕获序列中不同时间间隔之间的动态时间相关性。As shown in Figure 3, the convolutional attention module, which takes the original input sequence

Refinement to generate a new input sequence

, after the convolutional attention operation, increases the representation ability of the original input sequence; the spatial attention module, which can selectively capture the dynamic correlation between different input sequences; the gated recurrent unit, which can learn the input sequence The hidden layer representation of , and update the hidden state of the current moment according to the input sequence and its previous hidden state; temporal attention module. It can adaptively capture dynamic temporal correlations between different time intervals in a sequence.

Specific embodiments of the present invention will be introduced below from various embodiments:

Example 1

As shown in FIG. 1 , this embodiment provides a method for predicting missing values of multi-channel ocean observation time series scalar data.

Here we use the multi-channel ocean observation time series scalar dataset with missing chlorophyll values from the Canadian Ocean Network, which includes: depth, temperature, conductivity, salinity, oxygen content, dissolved oxygen concentration, Chlorophyll (including missing values), turbidity, pH value, wind speed and other ocean observation time series scalar data, the chlorophyll sequence with missing values is shown in Figure 7, where the x-axis represents the length of the chlorophyll sequence, and the y-axis represents the chlorophyll value. Missing values are shown as circles. Missing values in this dataset were imputed using the fixed value of 999. Combined with this data set, the technical solution of this embodiment is: prediction of missing values of multi-channel ocean observation time series scalar data based on a three-stage attention recurrent neural network prediction model, as shown in Figure 2, including the following steps:

(1) Take the dataset as the input of the model, and first perform data preprocessing on it to obtain the initial sequence. The preprocessing stage includes:

(1-1) The chlorophyll data to be filled is processed by linear interpolation to obtain the initial data;

(1-2) Construct a sample set without missing values, input the sample set without missing values into the model for training, and use the loss function to calculate the corresponding value.

，其中n表示不同类型序列的个数，

表示输入序列长度大小，

表示深度、风速、氧含量、溶解氧、浊度、温度、盐分这七个序列构成的多通道序列。(2) The chlorophyll sequence is used as the target sequence we need to predict, and the correlation between other sequences and the target sequence is measured by the Pearson correlation coefficient. By calculating the quotient of the covariance and standard deviation between the target series and the series of depth, temperature, conductivity, salinity, oxygen content, dissolved oxygen concentration, chlorophyll (with missing values), turbidity, pH value, wind speed, etc., we Select the seven sequences of depth, wind speed, oxygen content, dissolved oxygen, turbidity, temperature, and salinity, which are most related to the chlorophyll sequence and form the input sequence together with the chlorophyll sequence.

, where n represents the number of different types of sequences,

represents the length of the input sequence,

It represents a multi-channel sequence composed of seven sequences of depth, wind speed, oxygen content, dissolved oxygen, turbidity, temperature, and salinity.

和深度、风速、氧含量、溶解氧、浊度、温度、盐分这七个序列构成的多通道序列

。(3) Decompose the input sequence after (2) into chlorophyll sequences

A multi-channel sequence composed of seven sequences of depth, wind speed, oxygen content, dissolved oxygen, turbidity, temperature, and salinity

.

输入到CBAM模块中，CBAM模块如图4所示。首先通过平均池化和最大池化操作来聚合特征映射的空间信息，生成两个不同的空间上下文描述符分别表示平均池特征和最大池特征：

和

。(4) The multi-channel sequence

Input into the CBAM module, the CBAM module is shown in Figure 4. First, the spatial information of the feature map is aggregated through the average pooling and max pooling operations, and two different spatial context descriptors are generated to represent the average pooling feature and the max pooling feature respectively:

and

.

即：(5) As shown in Figure 5, these two descriptors are input into a shared network consisting of a multilayer perceptron and a hidden layer to generate a channel attention map

which is:

表示sigmoid函数，

和

表示多层感知机权重。in the formula

represents the sigmoid function,

and

Represents the multilayer perceptron weights.

与经过通道注意力映射的序列进行逐元素相乘操作。得到新的输入序列

，即：(6) Convert the original input sequence

Perform element-wise multiplication with the channel-attention-mapped sequence. get new input sequence

,which is:

表示逐元素相乘。In the formula,

Represents element-wise multiplication.

沿着通道轴应用平均池化和最大池化操作，通过两个池操作聚合特征映射的通道信息，生成两个空间上下文描述符：

和

。并将它们连接起来以生成有效地特征描述符，在连接的特征描述符上，我们应用卷积层去生成空间注意映射

，即：(7) As shown in Figure 6, the newly generated sequence is

Average pooling and max pooling operations are applied along the channel axis, and the channel information of the feature map is aggregated by the two pooling operations to generate two spatial context descriptors:

and

. and concatenate them to generate effective feature descriptors. On the concatenated feature descriptors, we apply convolutional layers to generate spatial attention maps

,which is:

表示sigmiod激活函数，

表示滤波器大小为

的卷积运算。In the formula,

represents the sigmiod activation function,

Indicates that the filter size is

convolution operation.

，即:(8) Perform element-wise multiplication of the new input sequence obtained in (6) and the sequence mapped by spatial attention to obtain the final refined output

,which is:

，作为空间注意模块的输入，通过空间注意力机制生成新的输入序列

，即：(9) The refined output

, as the input of the spatial attention module, which generates a new input sequence through the spatial attention mechanism

,which is:

表示第k个输入序列

，

表示t时刻编码器隐状态的注意权重，对注意权重

进行SoftMax函数标准化处理得到

是t-1时刻编码器隐状态，

和

是需要学习的参数矩阵，

是衡量在t时刻的第k个输入特征重要性的注意权重。In the formula,

represents the kth input sequence

,

Represents the attention weight of the hidden state of the encoder at time t, and the attention weight

Standardize the SoftMax function to get

is the hidden state of the encoder at time t-1,

and

is the parameter matrix to be learned,

is the attention weight that measures the importance of the kth input feature at time t.

(10) We get the attention weights, we can update the input sequence and the encoder hidden state at time t, namely:

，即：(11) Input the hidden state of the decoder and encoder at time t-1 and the hidden state of the encoder at time t into the time attention module, and obtain the context vector through the time attention mechanism

,which is:

是需要学习的参数矩阵，

是

时刻解码器的隐状态，

是t-1时刻编码器的隐状态，

是t时刻编码器的隐状态，

表示t时刻解码器的注意权重，对注意权重

进行SoftMax函数标准化处理得到

衡量在t时刻的第i个输入特征重要性的注意权重，

是上下文向量。In the formula,

is the parameter matrix to be learned,

Yes

the hidden state of the decoder at time,

is the hidden state of the encoder at time t-1,

is the hidden state of the encoder at time t,

Represents the attention weight of the decoder at time t, and the attention weight

Standardize the SoftMax function to get

The attention weight that measures the importance of the ith input feature at time t,

is the context vector.

，将它们与目标时间序列结合起来，并更新t时刻解码器隐状态

，即：(12) When the context vector at time t is obtained

, combine them with the target time series, and update the decoder hidden state at time t

,which is:

和b是将连接映射到解码器输入的参数矩阵，

是t-1时刻解码器的输入，

是计算出的上下文向量，

表示连接操作，

是经过线性变换后的新的输入，

是t-1时刻解码器的隐状态。In the formula,

and b is the parameter matrix that maps connections to decoder inputs,

is the input of the decoder at time t-1,

is the computed context vector,

represents the connection operation,

is the new input after linear transformation,

is the hidden state of the decoder at time t-1.

与T时刻解码器的隐状态

连接起来成为新的解码器的隐状态，从中做出最终预测：(13) Finally, the context vector

with the hidden state of the decoder at time T

The concatenation becomes the hidden state of the new decoder, from which the final prediction is made:

和向量

映射连接

，最终我们使用线性变化（

和

）生成最终的叶绿素预测结果。预测效果图如图8所示：In the formula, the matrix

and vector

map connection

, and finally we use a linear variation (

and

) to generate the final chlorophyll prediction result. The prediction effect diagram is shown in Figure 8:

(14) Fill the predicted chlorophyll data into the data set with missing chlorophyll values to obtain the final filling result. The result is shown in Figure 9, where the x-axis represents the length of the chlorophyll sequence, and the y-axis represents the value of the chlorophyll concentration , the circled part represents the value after the missing value is filled.

Here we take a part of the chlorophyll sequences with missing values of length 50, as shown in Figure 10, where the x-axis represents the length of the chlorophyll sequence, and the y-axis represents the chlorophyll value. The circled part represents the missing value of the chlorophyll sequence, where the missing value is represented by a fixed value of 999.

The result of the chlorophyll sequence after linear interpolation is shown in Figure 11, where the x-axis represents the length of the chlorophyll sequence, and the y-axis represents the value of the chlorophyll. The circled part represents the result of filling in missing values of chlorophyll sequences after linear interpolation.

The results of the chlorophyll sequence predicted by the model are shown in Figure 12, where the x-axis represents the length of the chlorophyll sequence, and the y-axis represents the value of the chlorophyll. The circled part represents the result of filling in the missing values of the chlorophyll sequence after model prediction.

Comparing Figures 10, 11, and 12, we can see that the accuracy of the recurrent neural network model based on three-stage attention for filling missing values is higher than that of linear interpolation.

This embodiment includes the following advantages:

(1) This embodiment predicts its current value based on the previous value of the chlorophyll sequence and the current and past values of the depth, wind speed, oxygen content, dissolved oxygen, turbidity, temperature, and salinity sequence, making up for the current missing value filling technology only. Defects that can be filled in for ocean single-channel observation time series scalar datasets.

(2) In this embodiment, the spatial attention module is used to replace the original input attention module, which can selectively capture the dynamic spatial correlation between different input sequences, so that the model can focus on the features associated with the prediction task in a targeted manner. The prediction accuracy of the model is improved, the training cost of the model is reduced, and the accuracy of the model for filling missing values is improved.

(3) This embodiment uses the convolution attention module to refine the input sequence. Compared with the original input attention module of DA-RNN, it can refine the input sequence and enhance the representation ability of the input sequence. It overcomes the gradient decay problem of the model in training large batches of data, and the prediction performance will not decrease due to the increase in the amount of data, and the prediction accuracy has good stability. The model is able to effectively impute large datasets with missing values.

Embodiment 2

This embodiment provides a multi-channel ocean observation time series scalar data missing value prediction system.

The technical solution of this embodiment includes the following modules:

1. Acquisition and preprocessing modules

Obtain a multi-channel ocean observation time series scalar dataset with missing chlorophyll values, and preprocess the dataset. The preprocessing process is as follows:

(1) Use linear interpolation to process the missing chlorophyll sequence, construct a sample set with no missing chlorophyll values, input the sample set without missing values into the model in the present invention for training, and use a loss function to calculate the corresponding value.

(2) Take the chlorophyll sequence as the target sequence we need to predict, and use the Pearson correlation coefficient to measure the depth, temperature, conductivity, salinity, oxygen content, dissolved oxygen concentration, chlorophyll (including missing values) in the marine multi-channel dataset. , turbidity, pH, wind speed and other sequences and the correlation between the chlorophyll sequence. By calculating the quotient of the covariance and standard deviation between the target sequence and other sequences, we select the sequence and the target sequence that are most related to the chlorophyll sequence of the seven sequences of depth, wind speed, oxygen content, dissolved oxygen, turbidity, temperature, and salinity Together they form the input sequence:

。

.

和由深度、风速、氧含量、溶解氧、浊度、温度、盐分组成的新的输入序列

，其中n表示新输入序列中不同类型序列的个数，

表示输入序列长度。(3) Decompose the data after (2) into chlorophyll sequences

and a new input sequence consisting of depth, wind speed, oxygen content, dissolved oxygen, turbidity, temperature, salinity

, where n represents the number of different types of sequences in the new input sequence,

Indicates the input sequence length.

2. Convolution attention module

作为输入，卷积注意模块（CBAM）依次推断出一个一维通道注意力映射

和二维空间注意力映射

。其总过程可以表示如下：Will

As input, the Convolutional Attention Module (CBAM) in turn infers a one-dimensional channel attention map

and 2D spatial attention maps

. The overall process can be expressed as follows:

表示逐元素相乘，在乘法过程中通道注意值沿着空间维度传播，

是最终细化的输出。in,

represents element-wise multiplication, during which the channel attention value is propagated along the spatial dimension,

is the final refined output.

和

，然后这两个描述符发送到一个共享网络中生成通道注意力映射

，共享网络由多层感知机和一个隐藏层组成，将共享层应用于每个描述符号后，我们使用元素求和合并输出特征向量，通道注意力计算公式如下：The specific calculation process is as follows. First, the spatial information of the feature map is aggregated through the average pooling and maximum pooling operations, and two different spatial context descriptors are generated to represent the average pooling feature and the maximum pooling feature:

and

, then these two descriptors are sent to a shared network to generate a channel attention map

, the shared network consists of a multilayer perceptron and a hidden layer. After applying the shared layer to each descriptor, we use element-wise summation to combine the output feature vectors. The channel attention calculation formula is as follows:

表示sigmoid函数，

和

表示多层感知机权重。in,

represents the sigmoid function,

and

Represents the multilayer perceptron weights.

，通过两个池操作聚合特征映射的通道信息，生成两个空间上下文描述符：

和

，空间注意力的计算如下：Computational spatial attention, we first apply average pooling and max pooling operations along the channel axis and concatenate them to generate efficient feature descriptors. Applying pooling operations along the channel axis can effectively highlight areas of information. On the concatenated feature descriptors, we apply convolutional layers to generate spatial attention maps

, which aggregates the channel information of feature maps through two pooling operations to generate two spatial context descriptors:

and

, the spatial attention is calculated as follows:

表示sigmiod激活函数，

表示滤波器大小为

的卷积运算。通过卷积注意机制对输入特征进行预处理，细化了输入特征，增强了输入特征的表征能力。in,

represents the sigmiod activation function,

Indicates that the filter size is

convolution operation. The input features are preprocessed through the convolutional attention mechanism, which refines the input features and enhances the representation ability of the input features.

3. Spatial attention module

，作为空间注意模块的输入，通过空间注意力机制生成新的输入序列

，即：the refined output

, as the input of the spatial attention module, which generates a new input sequence through the spatial attention mechanism

,which is:

表示第k个输入序列

，

表示t时刻编码器隐状态的注意权重，对注意权重

进行SoftMax函数标准化处理得到

是t-1时刻编码器隐状态，

和

是需要学习的参数矩阵，

是衡量在t时刻的第k个输入特征重要性的注意权重。通过空间注意机制，使得模型能够选择性地捕获不同输入特征之间的动态空间相关性。In the formula,

represents the kth input sequence

,

Represents the attention weight of the hidden state of the encoder at time t, and the attention weight

Standardize the SoftMax function to get

is the hidden state of the encoder at time t-1,

and

is the parameter matrix to be learned,

is the attention weight that measures the importance of the kth input feature at time t. Through the spatial attention mechanism, the model can selectively capture the dynamic spatial correlation between different input features.

4. Encoder

，编码器用于学习从

到

（在时间t）的映射：The encoder is essentially an RNN that, in machine translation, encodes the input sequence into feature representations. For the input sequence after spatial attention operation

, the encoder is used to learn from

arrive

Mapping (at time t):

表示编码器在t时刻的隐状态，m表示隐状态的大小，

表示一个非线性映射函数，这里我们使用门控循环单元（GRU）作为

来捕获序列中的长期依赖。GRU由2个门组成：重置门

，更新门

。GRU的更新过程如下所示：in,

represents the hidden state of the encoder at time t, m represents the size of the hidden state,

represents a nonlinear mapping function, here we use a gated recurrent unit (GRU) as

to capture long-term dependencies in sequences. GRU consists of 2 gates: reset gate

, update gate

. The update process of GRU is as follows:

为t-1时刻的编码器隐状态

和当前t时刻的输入

的连接，

为需要学习的参数。

表示sigmoid激活函数，

表示逐元素相乘。in,

is the hidden state of the encoder at time t-1

and the current input at time t

Connection,

parameters to be learned.

represents the sigmoid activation function,

Represents element-wise multiplication.

5. Time attention module

，在t时刻每个解码器隐状态的注意权重定义如下：The temporal attention mechanism is used in the decoding stage to model the dynamic temporal correlation between different time intervals in the input sequence, and the hidden states of the decoder and encoder at time t-1 and the hidden state of the encoder at time t are input to the temporal attention module , through the temporal attention mechanism, the context vector is obtained

, the attention weight of each decoder hidden state at time t is defined as follows:

是需要学习的参数矩阵，

是

时刻解码器的隐状态，

是t-1时刻编码器的隐状态，

是t时刻编码器的隐状态，

表示t时刻解码器的注意权重，对注意权重

进行SoftMax函数标准化处理得到

衡量在t时刻的第i个输入特征重要性的注意权重，

是上下文向量。in,

is the parameter matrix to be learned,

Yes

the hidden state of the decoder at time,

is the hidden state of the encoder at time t-1,

is the hidden state of the encoder at time t,

Represents the attention weight of the decoder at time t, and the attention weight

Standardize the SoftMax function to get

The attention weight that measures the importance of the ith input feature at time t,

is the context vector.

6. Decoder

，我们将它们与目标时间序列结合起来，并更t时刻的解码器新的隐状态

。When the context vector at time t is obtained

, we combine them with the target time series and update the new hidden state of the decoder at time t

.

和b是将连接映射到解码器输入的参数矩阵，

是t-1时刻解码器的输入，

是计算出的上下文向量，

表示连接操作，

是经过线性变换后的新的输入，

是t-1时刻解码器的隐状态。我们将上下文向量

与隐状态

连接起来成为新的解码器的隐状态，从中做出最终预测：

and b is the parameter matrix that maps connections to decoder inputs,

is the input of the decoder at time t-1,

is the computed context vector,

represents the connection operation,

is the new input after linear transformation,

is the hidden state of the decoder at time t-1. We will context vector

with hidden state

The concatenation becomes the hidden state of the new decoder, from which the final prediction is made:

和向量

映射连接

，最终我们使用线性变化（

和

）生成最终的叶绿素预测结果。Among them, the matrix

and vector

map connection

, and finally we use a linear variation (

and

) to generate the final chlorophyll prediction result.

7. Model validation

As shown in Figure 11, when the prediction result is obtained, the mean square error is used to calculate the loss value between the prediction result and the real value of the multi-channel data set after interpolation, and the network parameters of the model are adjusted to obtain the final Chlorophyll prediction results.

8. Missing value imputation

Fill the final chlorophyll prediction result into the missing value unit of the multi-channel dataset to get the filling result.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (10)

1. A method for predicting missing values of multi-channel ocean observation time series scalar data, characterized in that, comprising: Obtain ocean observation time series scalar data with ocean missing values; Based on the ocean observation time series scalar data, the TA-RNN model is used to obtain the ocean missing value prediction result; The TA-RNN model includes a convolutional attention module, a spatial attention module and a temporal attention module, the convolutional attention module is used to refine the ocean observation time series scalar data; the spatial attention module is used to capture the refinement The dynamic spatial correlation of the ocean observation time series scalar data; the time attention module is used to capture the dynamic time correlation between different time intervals in the output data of the spatial attention module. 2 . The method for predicting missing values of multi-channel ocean observation time series scalar data according to claim 1 , wherein after acquiring the ocean observation time series scalar data with ocean missing values, the method comprises: The ocean observation time series scalar data of the values are preprocessed to obtain the initial sequence. 3. The method for predicting missing values of multi-channel ocean observation time series scalar data according to claim 1, characterized in that, before said adopting the TA-RNN model, it comprises: if the ocean observation time series scalar data with ocean missing values is chlorophyll Sequence, select depth sequence, wind speed sequence, oxygen content sequence, dissolved oxygen sequence, turbidity sequence, temperature sequence and salinity sequence, according to depth sequence, wind speed sequence, oxygen content sequence, dissolved oxygen sequence, turbidity sequence, temperature sequence, The salt sequence and the chlorophyll sequence, construct a multi-channel sequence. 4. The method for predicting missing values of multi-channel ocean observation time series scalar data according to claim 3, wherein, according to the multi-channel sequence, a convolution attention module is adopted to obtain a channel attention map and a spatial attention map; The sequence of the channel attention map of the multi-channel sequence is multiplied element by element with the multi-channel sequence to obtain an initial refinement sequence; the sequence of the initial refinement sequence and the sequence of the spatial attention map of the initial refinement sequence are multiplied element by element , to get the final refined sequence. 5. The method for predicting missing values of multi-channel ocean observation time series scalar data according to claim 4, characterized in that, based on the final refinement sequence, a spatial attention module is used to capture differences between different input features in the final refinement sequence. Dynamic spatial correlation to get the input sequence. 6. The method for predicting missing values of multi-channel ocean observation time series scalar data according to claim 5, wherein, according to the input sequence, an encoder is used to learn the mapping from the input sequence to the hidden state of the encoder at time t , get the hidden state of the encoder at time t. 7. The method for predicting missing values of multi-channel ocean observation time series scalar data according to claim 6, characterized in that, according to the hidden state of the encoder at time t, a time attention module is adopted to capture the hidden state of the encoder at time t. In the sequence, the dynamic time correlation between different time intervals; the specific process of using the time attention module includes: According to the hidden state of the encoder at time t and the hidden state of the decoder at time t-1, determine the attention weight of each input feature at time t; The attention weight of each input feature to the predicted value; based on the attention weight of all input features to the predicted value at time t and the hidden state of the encoder at time t, the weighted sum of all the hidden states of the encoder is obtained, that is, the context vector. 8. The method for predicting missing values of multi-channel ocean observation time series scalar data according to claim 7, characterized in that, it is determined that the context vector at time t is combined with the target sequence at time t-1, and the decoder is updated at time t. the hidden state. 9. The method for predicting missing values of multi-channel ocean observation time series scalar data according to claim 8, wherein the context vector at time T is connected with the hidden state of the updated decoder at time T to become a new The hidden state of the decoder, predicting missing chlorophyll sequences. 10. A multi-channel ocean observation time series scalar data missing value prediction system, characterized in that it comprises: a data acquisition module, which is configured to: acquire ocean observation time series scalar data with ocean missing values; A prediction module, which is configured to: based on the ocean observation time series scalar data, using a TA-RNN model to obtain a prediction result of ocean missing values; The TA-RNN model includes a convolutional attention module, a spatial attention module and a temporal attention module, the convolutional attention module is used to refine the ocean observation time series scalar data; the spatial attention module is used to capture the refinement The dynamic spatial correlation of the ocean observation time series scalar data; the time attention module is used to capture the dynamic time correlation between different time intervals in the output data of the spatial attention module.

Images